Open-Weights AI Agents Hit the Mainstream — Kimi K3, Inkling, and GLM-5.2 Lead the Week

Yesterday we covered predictable billing on the open-weight side. Today the story expands: three major open-weights drops landed in the last 48 hours and they are already shipping into real agent harnesses.

Kimi K3 from Moonshot AI crossed the 2.8-trillion-parameter line with a July 27 open-weight promise. Thinking Machines Lab (Mira Murati’s outfit) released Inkling — a 975B MoE (41B active) multimodal model under Apache 2.0. GLM-5.2 from Zhipu AI is seeing production agent swaps that cut token spend by 70% while matching closed-model output on narrow workflows.

The Open-Weights Wave

The pattern is no longer “wait for the next GPT.” It’s “download the weights, route the boring loops here, keep frontier for the hard refactors.” Simon Willison’s coverage and the Thinking Machines announcement both highlight the same bet: customization beats one-size-fits-all when you control the stack.

Inkling ships with immediate fine-tuning on Tinker and hosted inference on Together, Fireworks, Modal, Databricks, and Baseten. Weights are live on Hugging Face. The lab is explicit that it is not chasing raw leaderboard supremacy — it is betting enterprise teams want something they can actually own and adapt.

Kimi K3’s open-weight timeline (promised by July 27) lines up with the same week GLM-5.2 showed up in Gumloop and other agent platforms as a drop-in replacement for Opus-class models. Early adopters report identical output on growth dashboards at a fraction of the cost.

Agent Frameworks Are Catching Up

The Q2 2026 framework releases (Microsoft Agent Framework 1.0, LangGraph per-node timeouts, CrewAI pluggable memory) now have viable open-weight backends that actually run the full loop locally or on self-hosted hardware. The MCP 2026-07-28 spec release candidate and A2A protocol adoption mean agent-to-agent handoff is becoming a first-class concern — and open weights remove the export-control friction that closed models still carry.

Yesterday’s Valkyrie story and today’s Inkling/Kimi/GLM cluster are two sides of the same coin: the tooling layer is maturing faster than the frontier model pricing.

The Spaghetti Takeaway

For teams already running harnesses, the move is obvious: route predictable, high-volume work to the open-weight stack (GLM-5.2 or Inkling on local or Fireworks) and reserve the closed frontier models for the genuinely novel refactors. The cost delta is now measurable in real deployments, not just benchmarks.

The week that started with predictable-cost agents ends with three heavyweight open-weights models that actually ship. The black box is cracking open.

Primary sources: Kimi K3 announcement and open-weight timeline, Thinking Machines Inkling release, GLM-5.2 agent cost savings reports, Simon Willison coverage of open-weights releases, Fireworks open-source LLM roundup.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.